We investigate how training languages should be selected for cross-lingual IPA ASR on unseen languages. Using Common Voice audio and Vox Communis phonetic transcripts, we train multilingual IPA-based ASR models for Upper Sorbian, Luganda, and Tatar under three linguistically motivated selection strategies: genealogical relatedness, geographic proximity, and phonological inventory overlap. We compare these strategies to a random baseline and evaluate performance with phone error rate. Linguistically informed selection generally improves transfer, but no single strategy is consistently optimal. Geographic proximity performs best for Luganda, phonological overlap is slightly best for Tatar, and none of the proposed strategies outperform random selection for Upper Sorbian. The results suggest that linguistic similarity aids low-resource ASR transfer, but that the most useful dimension of similarity varies by target language.

The use of ASR models in endangered language documentation has grown in popularity given the bottleneck of manual speech transcription. Meta’s Massively Multilingual Speech (MMS) model is particularly popular for its extensibility to low-resource languages. However, it is mostly trained on read speech data from the Bible, meaning it may not perform well on other domains. We evaluated this model on data collected as part of a larger language documentation and revitalization project focused on Akuzipik, a polysynthetic Alaska Native language. We also finetuned and evaluated the model on a small (1h) collection of speech. The original model performed well on a dataset that roughly matched the Bible training data in domain and writing style but struggled on a separate collection of spontaneous speech. Performance on spontaneous speech improved after finetuning on a sample of our full dataset, and error rates reduced less dramatically after finetuning only on read speech. Both finetuning scenarios show promise for future model improvement, especially considering the relative ease of collecting read speech data. This experiment confirms the challenge of transcribing spontaneous speech with the MMS ASR model but provides hope for improving model performance for language documentation purposes, even with scarce data.